# Read a scanned, photographed or handwritten page accurately

> Read a scanned, photographed or handwritten page accurately: 2 companies fully do it, 14 with a limit, 32 carry the tag. Every listing in Document parsing & extraction with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/document-extraction.scan-ocr/
- Markdown: https://www.anchorterminal.com/niches/document-extraction.scan-ocr/index.md (~1,850 tokens)
- Slim: https://www.anchorterminal.com/niches/document-extraction.scan-ocr/index.min.md (~430 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/document-extraction.scan-ocr/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 2 | 14 | 32 | 16 | 10 |

## What counts as done

Text of every line on the page in reading order, with no invented sentences and no dropped lines, the language named, and low-confidence words flagged, in one call per page or one job per file.

Measured, for our own tests: character error rate on a fixed set of faded, skewed and handwritten pages.

How agents ask for it: OCR this scan; what does this handwritten form say; turn these photographed pages into text; read this old typed document

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Azure Document Intelligence (Microsoft Corporation)**: The read model extracts printed and handwritten text with per-word confidence, lists handwriting languages and detects and returns language codes; acc-11 read a skewed page and a handwritten signature.
- **Google Cloud Document AI (Google LLC)**: Enterprise Document OCR reads printed and handwritten text in more than 200 languages with a quality assessment, and cells and tokens carry detectedLanguages with confidence; acc-15 and acc-17 found its OCR beat other tools on photos and long documents.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Adobe PDF Services / PDF Extract API (Adobe Inc.)**: Extract reads scanned PDFs, but the docs say it is optimised for English, skew and under-200 DPI lower quality, and the output schema has no word confidence or language field.
- **Amazon Textract (Amazon Web Services, Inc.)**: WORD and LINE blocks carry Confidence and TextType HANDWRITING or PRINTED, but handwriting and DetectDocumentText are English only (Category 1), no language is named in the response, and acc-08 found Textract did poorly where Mistral OCR worked.
- **Datalab (Endless Labs, Inc.)**: accurate mode targets scanned documents and word_bboxes adds a confidence per word, but no language field or handwriting statement was found in the conversion reference.
- **Docling (Docling Project (LF AI & Data Foundation))**: Several OCR engines (Tesseract with language detection, EasyOCR, RapidOCR, Surya, ocrmac), but confidence is a page-level grade, not flagged words, and acc-45 found the default Granite Vision preset gave malformed or empty output.
- **Extend API + MCP (CrowdView Inc. dba Extend)**: parse_performance handles handwriting and multilingual pages, languageDetected names languages, and each chunk carries minOcrConfidence and avgOcrConfidence, but low-confidence words are not flagged individually in the response read.
- **LandingAI Agentic Document Extraction (LandingAI, Inc.)**: Pages report status ok or failed and DPT-3 Verity gives per-word atomic grounding, but the parse response has no confidence field or language field, so low-confidence words are not flagged.
- **LlamaParse API + MCP (LlamaIndex, Inc.)**: OCR languages are a parameter, camera_photo_correction flattens phone photos and confidence_score_effort scores parsing quality per page, but low-confidence words are not flagged in what was read.
- **Mindee API (Mindee SAS (Mindee Inc. outside the Eurozone))**: A Raw Text OCR model returns the text of each page, but no word confidence, language field or handwriting statement was found in the pages read.
- **Mistral OCR API (Mistral AI (RCS Paris 952 418 325))**: Over 40 languages and confidence at page, block or word granularity, and acc-20 found it highly accurate, but acc-05 saw OCR 4.0 invent sentences that only a second model caught.
- **Nanonets API + MCP (Nano Net Technologies Inc.)**: OCR models (nanonets-ocr-s, gemini) with word bounding boxes and per-page confidence 0 to 100, but no language field or handwriting statement in the reference.
- **Nemotron 3.5 Parse (NVIDIA Corporation)**: Works on scans and reports OCR across 163 language and script entries, but the card says those results do not cover handwriting or degraded images, and output carries no confidence.
- **OpenDocRouter (LlamaIndex, Inc.)**: Takes images and scanned PDFs and reports per-page ok or error, but quality depends on the chosen model, and there is no word confidence or language field; layout elements carry a confidence only.
- **Reducto API + MCP (Reducto, Inc.)**: Blocks carry a confidence level and parse_confidence, return_citations gives a box per word, and a cookbook covers 60+ languages with detection, but no per-word confidence or language field in the response was seen.
- **Unstructured API + MCP (Unstructured Technologies, Inc.)**: Elements carry parse_confidence (agreement with OCR, 0 to 1, not a probability) and warnings flag degraded pages, but no language field or handwriting support was documented in the pages read.
- **Xiaomi-OCR-0 (Xiaomi Corporation)**: The card claims robustness on scanned, warped, photographed and skewed pages and shows handwriting examples, but output has no confidence or language field, and the benchmark figures are Xiaomi's own runs.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Veryfi API + MCP (Veryfi, Inc.)**: docs.veryfi.com was not read, since the Starter API terms bar scraping the documentation; the SDK README does not describe a full-text OCR output.

- not this job: Clairic (Cloudflare, Inc.): Images are accepted only as context for named fields; there is no full-text transcription output.
- not this job: Invofox (Invofox, Inc. (Delaware), with a registered branch in Spain): OCR feeds field extraction; the API returns no full-page transcription.

6 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- model invents plausible sentences with nothing to flag them
- embedded OCR text layer dropped with no warning
- handwriting or the script not supported (Textract handwriting is English only)

## Around it

Before: file-storage.agent-upload, mailbox-access.find-and-read. After: document-extraction.fields-from-business-docs, document-extraction.rag-ingest.

## Where this comes from

Weight 16 of 100 in Document parsing & extraction: 10 of 46 accounts; Mistral OCR and Google Document AI threads on Hacker News, with two accounts of hallucinated or silently dropped text

As data: https://www.anchorterminal.com/api/v1/use-cases/document-extraction.scan-ocr.json, or the MCP tool get_use_case with id document-extraction.scan-ocr. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
